|
Received: 22 July 2016 Accepted: 5 October 2017 DOI: 10.1111/desc.12634
PA P E R
Children retain implicitly learned phonological sequences better than adults: a longitudinal study Eleonore H.M. Smalle1,2 | Mike P.A. Page3 | Wouter Duyck4 | Martin Edwards1,2 | Arnaud Szmalec1,2,4 1 Psychological Sciences Research Institute, Université catholique de Louvain, Louvain-la-Neuve, Belgium 2
Institute of Neuroscience, Université catholique de Louvain, Louvain-la-Neuve, Belgium 3
Department of Psychology, University of Hertfordshire, Hatfield, UK 4
Department of Experimental Psychology, Ghent University, Ghent, Belgium Correspondence Eleonore Smalle, Université catholique de Louvain, Psychological Sciences Research Institute, Institute of Neuroscience, Place Cardinal Mercier 10, room C314, B-1348 Louvain-la-Neuve, Belgium. Email:
[email protected] Funding information Fonds de la Recherche Scientifique – FRS – FNRS, Belgium
Abstract Whereas adults often rely on explicit memory, children appear to excel in implicit memory, which plays an important role in the acquisition of various cognitive skills, such as those involved in language. The current study aimed to test the assertion of an age- dependent shift in implicit versus explicit learning within a theoretical framework that explains the link between implicit sequence memory and word-form acquisition, using the Hebb repetition paradigm. We conducted a one-year, multiple-session longitudinal study in which we presented auditory sequences of syllables, co-presented with pictures of aliens, for immediate serial recall by a group of children (8–9 years) and by an adult group. The repetition of one Hebb sequence was explicitly announced, while the repetition of another Hebb sequence was unannounced and, therefore, implicit. Despite their overall inferior recall performance, the children showed better offline retention of the implicit Hebb sequence, compared with adults who showed a significant decrement across the delays. Adults had gained more explicit knowledge of the implicit sequence than children, but this could not explain the age-dependent decline in the delayed memory for it. There was no significant age-effect for delayed memory of the explicit Hebb sequence, with both age groups showing retention. Overall performance by adults was positively correlated with measures of post-learning awareness. Performance by children was positively correlated with vocabulary knowledge. We conclude that children outperform adults in the retention over time of implicitly learned phonological sequences that will gradually consolidate into novel word-forms. The findings are discussed in the light of maturational differences for implicit versus explicit memory systems that also play a role in language acquisition. A video abstract of this article can be viewed at: https://youtu.be/G5nOfJB72t4
RESEARCH HIGHLIGHTS
1 | INTRODUCTION
• Children are very good at memorizing implicitly learnt word-forms.
Children appear to be superior in skill learning and in the acquisition
• Despite their higher short-term memory capacity, adults are more
of the implicit, procedural knowledge that underlies formal aspects of language (Johnson & Newport, 1989; Newport, Bavelier, & Neville,
prone to forgetting. • The child advantage remains noticeable up to a consolidation pe-
hand, are argued to be superior on most tests of cognitive abilities, es-
riod of one year. • The findings suggest developmental changes in implicit memory for novel word-forms. Developmental Science. 2017;e12634. https://doi.org/10.1111/desc.12634
2001; Ullman, 2004; Weber-Fox & Neville, 1996). Adults, on the other pecially those that rely on explicit, declarative knowledge or attention (Craik & Bialystok, 2006; Murphy, McKone, & Slee, 2003). Recently,
wileyonlinelibrary.com/journal/desc
© 2017 John Wiley & Sons Ltd | 1 of 15
|
SMALLE et al.
2 of 15
Janacsek and colleagues compared nine different age cohorts (from 4
question of whether the decline between sessions observed in adults
to 85 years old) on the serial reaction time (SRT) task (Janacsek, Fiser,
might be due to a ceiling effect at the end of the first session, allowing
& Nemeth, 2012). The SRT is a well-known implicit skill-learning task
more room for them to decline across sessions compared with chil-
in which participants are asked to press a button that corresponds to
dren. To answer this question, they compared children and adults who
a particular location on the screen. Unannounced to the participant, a
showed comparable scores on the last trial of session A. They found
particular sequence is repeated throughout the task, and results show
that these children still retained better across the one-hour delay than
better accuracy and faster response times for stimuli consistent with
corresponding adults. This shows that children retain nonwords better
its sequence structure than for inconsistent stimuli. This is taken to
than adults, even when initial learning performance is matched. This
indicate implicit learning of the sequence. Interestingly, the authors
finding supports the hypothesis of a maturational effect in delayed
observed a peak in task performance before the age of 12, after which
memory which, as the authors argue, might have implications for the
performance dropped significantly. In another version of the task,
existence of sensitive periods in language-learning.
however, where participants were explicitly made aware of the repeti-
Here, we aimed to investigate child–adult differences in delayed
tion, using instructions and visual cues, they found no age-dependent
memory for novel word-forms that were either acquired implicitly or
change in learning (Nemeth, Janacsek, & Fiser, 2013). Nonetheless, in
explicitly through repetition learning of syllable sequences. We con-
this latter case, additional measures of explicit knowledge concerning
ducted a one-year, four-session learning study with a group of chil-
the repeating sequence (using verbal reports) revealed a different age-
dren (8–9 years old) and a group of adults, who were presented with
effect, with significantly better scores on explicit knowledge of the
novel sequences of nine syllables for immediate serial recall. These se-
sequence after the age of 12.
quences were segmented into three chunks of three syllables by virtue
Based on these findings, Nemeth et al. (2013) argued that children’s
of the co-presentation of pictures of aliens, as described below. In the
superiority in various skills might have its basis in underlying implicit-
task we used, which is an adaptation of the well-known Hebb learning
learning abilities that shift to explicit-learning mechanisms later in life
task (Hebb, 1961), there were certain sequences of items (often called
(Poldrack & Packard, 2003). This might, in turn, be attributed to a neu-
Hebb sequences) that were repeated in exactly the same order every
rological change at the level of two separate long-term memory sys-
third trial. Recall performance for such Hebb sequences usually im-
tems, namely an early developed basal-ganglia system (the procedural
proves relative to other non-repeating (filler) sequences. This finding,
memory system), that is involved in implicit skill learning, and the late
known as Hebb Repetition Learning (HRL), is thought to reflect a grad-
maturing frontal/medial-temporal circuits (the declarative memory
ual transfer of serial-order information held in short-term memory to
system), that is assumed to rely on explicit attention processes (Perez,
a longer-lasting representation of the sequence in long-term memory.
Peynircioglu, & Blaxton, 1998; Robertson, 2012; Squire, 1992). With respect to human language, only a few studies have directly
It has been hypothesized that this HRL provides a source of evidence for the role of implicit sequence memory mechanisms in
compared implicit language-learning capacities between children and
language learning, particularly with respect to the long-term acquisi-
adults under equivalent learning conditions. Ferman and Karni (2010),
tion of novel phonological word-forms (Majerus & Boukebza, 2013;
for example, compared 8- and 12-year old children and adults when
Page & Norris, 2008, 2009; Szmalec, Duyck, Vandierendonck, Mata,
learning the same implicit artificial-grammar rule, and found better
& Page, 2009; Szmalec, Page, & Duyck, 2012). During novel word-
within-session gains in adults. An interesting observation is, how-
form acquisition, learners need to segment word units from pho-
ever, that children seem to have the capacity to outperform adults
nological (speech) input, either by tracking transitional probabilities
on the subsequent formation of long-term memory representations
among sounds in the incoming stream (Saffran, 2001; Saffran, Aslin,
(i.e., in an offline consolidation phase), irrespective of sleep (Adi-Japha,
& Newport, 1996), or (alternatively) by the application of chunking
Badir, Dorfberger, & Karni, 2014; Ashtamker & Karni, 2013; Ferman &
mechanisms in memory (Thiessen & Erickson, 2013). Regularly oc-
Karni, 2010; Wilhelm, Prehn-Kristensen, & Born, 2012). In the study
curring sub-sequences gradually consolidate into long-term memory
by Ferman and Karni (2010), reaction-time data revealed greater
due to repeated exposure of their serially ordered constituents (i.e.,
between-session gains, with respect to an implicitly learned grammar
syllables/phonemes) (Page & Norris, 2008, 2009). Hebb repetition
rule, for children than for adults, over a between-session delay of up
learning of, for instance, a sequence of consonant-vowel syllables
to two months. Similarly, Bishop and colleagues found that children
(e.g., lo-fo-du) is therefore, by hypothesis, functionally equivalent to
outperformed adults on the offline retention of novel word-forms
the long-term memorization (hence, lexicalization) of a correspond-
(Bishop, Barry, & Hardiman, 2012). This was found using a non-word
ing novel word-form (lofodu). In Page and Norris (2008, 2009), this
repetition task in which unfamiliar polysyllabic nonwords were pre-
learning is explained within a connectionist model that involves
sented for repetition (session A), and re-presented for repetition after
several layers of localist units representing phonemes or syllables,
a delay of one hour (session B). Overall, the scores of the children
together with a learning mechanism that progressively consoli-
tended to improve over the delay, while those of the adults declined
dates a short and repeatedly presented sequence into one (local-
significantly. No age differences were observed for learning within
ist) chunk representation in long-term memory. Experimental work
sessions, despite the observation that children initially started at a
has supported this idea by showing that sub-sequences acquired
lower baseline, owing to the fact that they were exposed to the same
through HRL are represented in the mental lexicon just like novel
learning material as adults. Importantly, the authors addressed the
words, exhibiting competition effects with existing words (which
|
3 of 15
SMALLE et al.
is an important indication of lexicalization; see Leach & Samuel,
but not, or less so, for the explicit (instructed) Hebb sequence at which
2007; Szmalec et al., 2009; Szmalec et al., 2012). HRL is also fast
adults are hypothesized to perform well. Moreover, there is reason to
and long-lasting (Page, Cumming, Norris, McNeil, & Hitch, 2013),
expect that the child advantage in the implicit condition should itself
and correlated with measures of word-learning and vocabulary
depend on the degree to which the children and the adults notice the
knowledge (Majerus & Boukebza, 2013; Majerus, Poncelet, Elsen,
unannounced repetition, and therefore switch to more explicit learn-
& Van der Linden, 2006); it is observed in children (Mosse & Jarrold,
ing. That is, the age-effect should be reduced for participants who are
2008; Smalle et al., 2016); and it has been shown to be impaired in
more aware of the repetition. Correlations with post-learning aware-
people with certain language disorders (Archibald & Joanisse, 2013;
ness of the unannounced Hebb sequence, vocabulary knowledge and
Bogaerts, Szmalec, De Maeyer, Page, & Duyck, 2016; Bogaerts,
short-term memory capacity were also investigated to explore the
Szmalec, Hachmann, Page, & Duyck, 2015; Hsu & Bishop, 2014;
potential role of explicit attention processes, lexical knowledge and short-term memory capacity in Hebb repetition learning (HRL) in both
Szmalec, Loncke, Page, & Duyck, 2011). Learning in the Hebb repetition task is further considered to
age groups.
depend on implicit learning mechanisms (although explicit awareness often naturally evolves across the repetitions) (Guerard, Saint- Aubin, Boucher, & Tremblay, 2011). For instance, neither explicit awareness of the repeating sequence, nor explicit reproduction of the sequence (i.e., overt recall) during the recall phase, nor focal
2 | METHODS 2.1 | Participants
hippocampal lesions, seem necessarily to affect learning (Couture &
Fifty-five participants took part in our longitudinal study, 29 chil-
Tremblay, 2006; Gagnon, Bedard, & Turcotte, 2005; Gagnon, Foster,
dren and 26 adults (Table 1). Children were recruited from two dif-
Turcotte, & Jongenelis, 2004; Guerard et al., 2011; Kalm & Norris,
ferent schools in and around Ghent (Belgium). Adults were recruited
2016).
by means of online advertising. All participants were Dutch speaking
In the present study, participants were made explicitly aware of
and had no reported history of speech, language, learning or hear-
only one of two Hebb sequences (i.e., the sequences and their co-
ing difficulties. All participants gave informed consent and received
presented pictures) by the use of intentional encoding instructions is-
financial compensation (€20) for their participation. Parental consent
sued prior to training (Gagnon et al., 2005), as well as by use of explicit
was obtained for the children’s participation. The study was approved
cuing during the task (similar to Nemeth, Janacsek, & Fiser, 2013). The
by the Psychology Ethics Commission at the Université catholique de
other Hebb sequence remained implicit, in the sense that its presence
Louvain.
was unannounced to participants. On the first day, two learning sessions were carried out, separated by a four-hour delay. One week and, then again, one year after learning, the sequences were re-presented
2.2 | Materials
to measure their long-term retention (cf. Bishop et al., 2012). Based
Sequences of nine syllables were presented to the participants for
on the previous observation of better delayed memory for nonwords
immediate serial recall. All syllables had a consonant-vowel structure
in children than in adults, we expected to see better retention of the
(CV). The length of the sequences was matched to an average adult
Hebb sequences in children than in adults, up to a one-year delay. In
digit-span level, with two more items added to avoid ceiling effects.
addition, in line with previous findings that showed a decline in im-
The same length was used for both the child and adult groups so that
plicit learning, but not in explicit learning, after age 12 (Janacsek et al.,
potential differences between the groups could not be attributed to
2012; Nemeth, Janacsek, & Fiser, 2013), we expected to see child ad-
differential input. To create the stimulus-sequences, four different
vantages particularly for the implicit (unannounced) Hebb sequence
sets (A to D), each set consisting of three CV-pools (1, 2, 3) with nine
T A B L E 1 Gender, laterality, age, cognitive test scores and session intervals divided by age group
Children (n = 15F/14M) Age (yr)
8.7 (.47)
Digit Forward Span (max = 9)
5.8 (1.5)
Peabody Vocabulary Test – 3rd ed. (PPVT-III), max = 204, rs
(a)
111 (12)
Adults (n = 17F/9M) 35.5 (11.09) 7.3 (1.7) (b)
185
(7)
Group difference p < .001 p < .001 p < .001
Session interval A – B (in hours) B – C (in days) C – D (in months)
4:47 (.03)
4:34 (.03)
ns
7 (1.1)
7 (1.0)
ns
11.9 (.30)
12.0 (.47)
ns
F = female; M = male; rs = raw score; (a) one missing value replaced by the mean; (b) four missing values replaced by the mean.
|
SMALLE et al.
4 of 15
different CVs in each CV-pool (see Table 2), were generated using
randomly. For this reason, there was no systematic mapping for the
WordGen (Duyck, Desmet, & Verbeke, 2004). We confirmed that the
filler sequences between CV-syllables and drawings.
summed bigram frequencies (as measured within the speaker’s native language; Duyck et al., 2004) for the entire sequences were compara-
2.3 | Procedure
ble within and across sets (i.e., all ps > .05). Each CV-pool within a set could be used either to generate filler sequences or, directly, as one
The experiment started with a familiarization phase during which
of the two Hebb sequences. This was counterbalanced across par-
participants were exposed to the syllable materials. Each participant
ticipants in a Latin square order (i.e., sequences 123, 231, 312 were
listened carefully to the CVs that were also presented visually on
used for Fillers, Hebb 1 and Hebb 2, respectively, for participants 1, 2
the screen. Participants were instructed to repeat the CVs out loud,
and 3), so that Hebb learning could not be confounded by unknown
and they were corrected by the experimenter if the CV was not pro-
particularities in the sequences themselves. Every three subsequent
nounced well. This familiarization phase was repeated three times to
participants were allocated to one of the four sets (A–D) in turn. The
ensure that participants clearly distinguished the different syllables.
allocation of participants to sets was matched across child and adult
All CVs were recorded by a female professional diction/pronuncia-
groups.
tion teacher and presented auditorily at 60 dB using Bose QC 15
To facilitate word-segmentation processes in the HRL task,1 and
headphones.
simultaneously to make the task more attractive for the children and
On the first day of testing, two sessions of 36 sequence trials were
similar to natural word-learning (where meaning is provided), the se-
presented for immediate serial recall: these were sessions A and B.
quences were presented together with drawings of aliens derived from
The two sessions were separated by a four-hour break in which par-
the TOTimals stimulus set (Gupta, 2003; Schwartz & Smith, 1997). For
ticipants continued their regular daily activities. The two types of
each sequence of nine syllables, a given drawing was presented to-
Hebb sequence were mixed within the same session and each one
gether with each subgroup of three successive syllables within that se-
was repeated on every third trial, giving 12 repetitions for each of the
quence (Figure 1), so that each three-syllable sub-sequence potentially
two Hebb sequences. These were interspersed with a total of 12 filler
became associated with the simultaneously presented drawing. From
sequences, resulting in 36 sequence trials per session. The syllables
a total of 36 different drawings, nine drawings were associated with
within a sequence trial were presented auditorily, one at a time, each
each set of three CV-pools. Of these nine drawings, three drawings
having a duration of 500 ms and a silent inter-stimulus interval of
were associated with the pool used to generate the filler sequences;
500 ms. The Hebb learning procedure was similar to that of previous
another three drawings were associated with the first Hebb sequence
studies (e.g., Smalle et al., 2016; Szmalec et al., 2009). Unlike previous
and the remaining three drawings were associated with the second
studies, we accompanied each sub-sequence of three successive sylla-
Hebb sequence. In each case, the three drawings associated with a
bles within a given sequence with the visual presentation of a picture
given type of sequence (Filler, Hebb 1, Hebb 2) were clearly distin-
of an alien. The participants were told that they would see “aliens from
guishable from each other and they were always presented together
the fictional planet Erinominus”, presented simultaneously with audi-
with the CV-sequence in the same order. The order of the CV-syllables
tory sequences of syllables. Each drawing appeared on the screen for
within the Hebb sequences remained fixed throughout learning, so
3 s, and thus matched the time during which the three corresponding
that, for the Hebb sequences, the mapping was constant between a
syllables were heard by the participant. The task was to repeat the
given drawing and a given sub-sequence of three successive syllables.
syllables in the same order as presented, after all nine syllables had
The order of the syllables within the filler sequences was determined
been heard.
Set A
B
C
D
CV-pool
CV1
CV2
CV3
CV4
CV5
CV6
CV7
CV8
CV9
1
fu
ra
wo
mu
zu
vo
za
ku
wu
2
wi
zi
vi
bo
ro
gi
ji
pe
hi
3
su
ke
fo
wa
mo
ru
gu
ki
ga
1
fo
ro
zu
wi
gu
di
zi
fi
fu
2
ru
ga
hi
ku
wu
ke
vi
fe
mo
3
gi
wo
pe
za
da
wa
ra
vo
bo
1
wo
da
vo
fu
gi
su
hi
ga
bo
2
fo
ra
pe
ku
wa
ki
wu
za
gu
3
ru
wi
mo
zu
di
vi
ji
ke
mu
1
mu
da
vo
di
wu
vi
fu
ve
wa
2
fo
su
gi
ke
ra
fe
ru
ji
wo
3
ga
ro
zu
hi
wi
gu
fi
mo
za
T A B L E 2 Stimulus materials for the Hebb learning task
|
5 of 15
SMALLE et al.
…
1
2
12
wi zi vi bo ro gi ji pe hi
…
1
2
12
…
Session D
12 months
2
Session C
7 days
1
Session B
4 hours
Session A
12
1
2
Filler recall ? ? ?
F I G U R E 1 Design of the experiment. The Hebb Repetition Learning Task (HRL). For session A, B and C, 36 trials were presented for immediate serial recall (12 filler sequences and two Hebb sequences that were each repeated 12 times). In session D, six trials were presented (two filler sequences and two Hebb repetitions)
fu ra wo mu zu vo za ku wu
Himplicit recall ?
? su ke fo wa mo ru gu ki ga ?
Hexplicit recall ? ?
?
Before learning started, the participants were also explicitly told
in sessions C and D. The same Hebb procedure was used as above,
that one sequence of nine speech sounds and the associated aliens
but without a practice session. Participants were presented with 12
were always repeated in the same order during the task. This is re-
repetitions (for Hebb sequences) or presentations (for fillers) of each
ferred to as the explicit Hebb sequence or Hexplicit in the remainder
sequence type in session C (i.e., 36 trials in total), and two more se-
of the paper. Every time Hexplicit was repeated, it was cued with a
quence repetitions in session D (i.e., six trials in total). Across all four
beeping tone and a red border around the screen. Participants were
sessions, 114 trials were presented, including 38 fillers, 38 explicit
not made aware of the occurrence of the second repeating sequence,
Hebb repetitions, and 38 implicit Hebb repetitions.
from now on referred to as the implicit Hebb sequence or Himplicit.
At the end of each of the first three sessions (i.e., A, B, and C),
Two filler sequences were presented to familiarize participants with
a verbal awareness report assessing explicit knowledge of the re-
the immediate-serial-recall procedure. Learning started with presenta-
peating sequences was administered. This questionnaire included
tion of a filler sequence followed by one of the two types of Hebb se-
three increasingly specific questions (based on Gagnon et al., 2005;
quence (implicit or explicit), the other type of Hebb sequence (explicit
Janacsek et al., 2012; Turcotte, Gagnon, & Poirier, 2005). In a first
or implicit), again a filler sequence, etc. In total, 12 filler sequences
question, the experimenter asked whether participants were aware
and 12 repetitions of the two types of Hebb sequence (i.e., 36 trials in
of the repeating Hebb sequence that followed the beep sound, pre-
total) were presented. The order of the two types of Hebb sequence
sented within a red border. This was asked to ensure that the partici-
was counterbalanced across participants. Immediately after presenta-
pant had understood the explicit instructions (this was the case for all
tion of each sequence, a recall screen was presented with the first
participants). In a second question, the experimenter asked whether
alien and a question mark signaling that the participant had to repeat
they noticed something else special in the task. Participants were free
the first three syllables of the sequence. Subsequently (initiated by a
to respond. If the participant answered (or gave the impression) that
manual press on the spacebar), the next alien appeared with a ques-
there was another repetition, the experimenter acknowledged the
tion mark signaling the requirement to recall the next three syllables
participant as being aware. After this, two sub-questions were pre-
of the sequence, and so on, until all nine CVs of the sequence were
sented to all participants (also the unaware participants) (1) to rate
recalled. Participants were allowed to say “blank” when they had for-
confidence (about what they had noticed) on a scale from 1 “Not
gotten a syllable at a particular serial position. The uttered syllables
certain at all” to 10 “100% certain” and (2) to determine when in the
were recorded and simultaneously written down by the experimenter.
task they started noticing something (i.e., at the beginning, middle or
For a syllable to be judged correct, the consonant and vowel in the
at the end of the task). Confidence ratings after each session were
heard syllable had to be present in the response, in the correct order,
used as measure of post-learning awareness throughout the remain-
and without the insertion of additional phonemes. Consonants and
der of the paper. Participants who did not notice anything special,
vowels that were pronounced ambiguously (e.g., /p/ for /b/, or /e:/ for
or noticed something irrelevant on the second question, were au-
/i:/) were always scored in favor of the presented phoneme provided
tomatically allocated the number zero on the confidence scale (and
that the ambiguous sound (i.e., /p/ or /e:/) did not occur elsewhere in
their answers on the sub-questions were ignored) so that no ex-
the sequence.
clusion was needed in the analyses. The use of verbal reports after
One week and, then again, one year after learning (see also
learning, in particular confidence ratings, is a common procedure
Table 1), participants returned to the lab for re-testing on the HRL task
and a relatively sensitive measure for demarcating explicit memory
|
SMALLE et al.
6 of 15
in humans (Cleeremans, Destrebecqz, & Boyer, 1998; Sandberg,
sequence of two or more items between the first error from the left and the first error from the right was counted. Finally, any other
Timmermans, Overgaard, & Cleeremans, 2010). Finally, we measured verbal short-term memory capacity using the
items that occurred in the correct position from left to right were
forward digit-span subscale of the Wechsler Intelligence Scales for
counted. The maximum possible score using this method was nine.
Children-Revised (WISC-R) (Wechsler, 1974) following session B, and
All trial scores were transformed into percentage scores. The trial
vocabulary knowledge using the Peabody Picture Vocabulary Test (3rd
scores at each session are plotted in Figure 2 (upper panel). Initial
Edition, PPVT-III) (Dunn & Dunn, 1981), which was administered at the
learning gains were determined by comparing the average scores
end of the experiment. Missing values were replaced by group means
over the first four trials with the average scores over the last four
(Table 1). In addition, immediately at the end of learning in session C,
trials of session A (Turcotte et al., 2005).2 This comparison is plot-
the 12 alien pictures that were presented during the Hebb task were
ted in Figure 2 (lower panel). Offline consolidation was determined
randomly presented on the screen. Participants were asked whether
by fitting within-session data with a power-law function model for
they could recall the names of the aliens they had seen during the ex-
both groups (Pan & Rickard, 2015; Rickard, 2007). Delayed retention
periment. This naming task had no a priori theoretical motivation and
was then tested by comparing trial performance in the subsequent
was added primarily to make the procedure more game-like and thus
session (i.e., session B, C, or D) with the performance level predicted
attractive for the participating children.
by an extrapolation of the performance in the previous session (i.e., session A, B, or C, respectively) (see also Adi-Japha et al., 2014, who used a similar approach). Extrapolation using power-law fitting is ar-
2.4 | Statistical analyses
gued to be a more reliable measure for offline consolidation effects
We scored trial performance using McKelvie’s scoring method
compared with measures based solely on pre-post difference scores
(McKelvie, 1987). In a first step, the number of items that were in the
(Pan and Rickard, 2015). All significant results are reported together
correct position from left to right, up to the first error, was counted.
with the ɳ2p effect size and Greenhouse Geisser (GG) correction
Second, the same step was repeated from right to left, again up
factors where applicable. Planned comparisons were conducted by
to the first error. After this, the number of items in any correct
Fisher’s LSD comparisons.
Filler Explicit Hebb Implicit Hebb
Children
Adults
100 Correct responses (%)
Correct responses (%)
100 80 60 40 20 0
Initial
0
Final
4
8
Session A
12 13
17 21 Trials Session B
25
29
33
37
80 60 40 20 0
Session C
Final
8
Session A
Session D
12 13
17
21
25
29
33
37
Session B
Session C
Session D
100
Correct responses (%)
Correct responses (%)
4
Trials
100 80 60 40 20 0
Initial
0
Initial
Final
80 60 40 20 0
Initial
Final
F I G U R E 2 Performance (percentage of correct scores) as a function of Sequence Type (Filler, Explicit Hebb, Implicit Hebb) and Age group (Children and Adults). 1. Upper panel: performance across all sessions (i.e., repetition 1–12 for session A, repetition 13–24 for session B, repetition 25–36 for session C and repetition 1–2 for session D). 2. Lower panel: Performance across two time points (initial vs. final) in session A. Error bars denote SEM
|
7 of 15
SMALLE et al.
3 | RESULTS
3.2 | Offline consolidation
3.1 | Initial Hebb learning
Based on previous observations in language and skill learning studies (Ashtamker & Karni, 2013; Bishop et al., 2012), we had anticipated
The average performance levels for the initial and for the final four
that children would retain the sequences better between sessions. To
trials of session A were entered into a mixed-measures analysis of
estimate delayed retention differences, a power-law function was fit-
variance with Age group (child vs. adult) as between-subjects fac-
ted to the data of the initially trained Hebb sequences, and we tested
tor, and Time point (initial vs. final) and Sequence type (Filler vs.
for the possibility of additional improvements or decrements beyond
Himplicit vs. Hexplicit) as repeated measures. While a main effect
what could have been expected from an extrapolation.
of Sequence type in favor of the Hebb sequence might provide some evidence of learning that sequence, only the demonstration
3.2.1 | Across four hours
of improvements in performance for repeated Hebb sequences relative to improvements in the baseline filler sequences may be
A power-law function, f(r) = arb where r is the repetition number of the
considered a pure indication of learning. Thus, an interaction be-
Hebb sequence and a and b are parameters estimated from the data,
tween Sequence type and Time point, due to higher scores on the
was fitted to the average recall data of session A for both the implicit
Hebb sequences for the final time point, would provide evidence
and explicit Hebb sequence. This yielded a significant fit for the two age
for Hebb learning (Smalle et al., 2016). Overall, adults showed
groups for both Hexplicit (R2 > .96; p < .001) and Himplicit (R2 > .93;
higher recall scores than children (main effect of Age group, F(1,
p < .001). The exponential coefficients (bs) were significantly different
53) = 32.1, p < .001, ɳ 2p = .38). There was also a significant main
from zero (ts > 14.69, p < .001 for Hexplicit; and ts > 10.19, p < .001 for
effect of Time point, F(1, 53) = 95.4, p < .001, ɳ2p = .64, such that
Himplicit), indicating significant learning in session A for both sequences
recall scores for the final time point were higher than recall scores
in both groups. The power-law models for the group-averaged data for
for the initial time point, and a significant main effect of Sequence
the two types of Hebb sequence were extrapolated to four additional
type, F(2, 106) = 51.5, p < .001, ɳ2p = .49. Comparisons revealed
trial points and the differences between this extrapolation and each in-
higher recall scores for both types of Hebb sequence compared
dividual’s performance at the four-hour post-learning session (i.e., ses-
with the filler sequence (Filler vs. Hexplicit, F(1, 53) = 103.4, p
.15, ɳ2p = .037.
compared between groups. For Himplicit, this revealed a main effect of
Furthermore, there was a significant interaction between Time
Age group, F(1, 53) = 5.7, p < .05, ɳ2p = .10 with children showing overall
point and Sequence Type, F(2, 106) = 41.3, p < .001, ɳ2p = .44.
larger gains across the delay (M difference score = 8.7%, SE = 4.1, t(28)
Planned comparisons of the two-way interaction revealed a signifi-
= 2.1, p < .05), in contrast with adults who showed numerical but non-
cant improvement across time points for Hexplicit, F(1, 53) = 90.83,
significant decrements across the delay (M difference score = −6.5%,
p < .001, ɳ2p = .63, and Himplicit, F(1, 53) = 73.5, p < .001, ɳ2p = .58,
SE = 4.9, t(25) = −1.3, p = .196). In a post-hoc analysis, to control for
but not for the filler sequence, F < 1. There were no interactions
unavoidably differing levels of overall recall across groups, we compared
with Age group, F < 1. Overall, these results show that there was
children and adults who showed similar recall scores on the last trial of
significant Hebb learning within session A, but its extent was not
session A. The 75% of children that showed the highest recall scores on
different for children and adults.
Hebb - Adults Extrapolated - Adults Hebb - Children Extrapolated- Children
F I G U R E 3 Offline consolidation performance after four hours as a function of Sequence Type (Explicit Hebb, Implicit Hebb) and Age group (Children and Adults). Hebb learning during session A is shown together with an extrapolation to four additional data points in session B. (1) Left panel: performance for the implicit Hebb sequence. (2) Right panel: performance for the explicit type of Hebb sequence. Error bars denote SEM
Correct responses (%)
Implicit
Explicit
100
100
80
80
60
60
40
40
20
20
0
0
0
2
4
6
8
10
12
14
16
Session A
0
2
4
6
8
10
12
14
16
Trials
Trials Session B
Session A
Session B
|
SMALLE et al.
8 of 15
trial 12 of session A (n = 21; M trial score = 60.3%, SD = 21.3) were com-
included for extrapolation. A power-law function, f(r) = arb, was
pared with the 75% of the adults that showed the lowest recall scores on
fitted to the average trial scores of session C for the two types
this trial (n = 19; M trial score = 61.4%, SD = 25.5), these groups having
of Hebb sequences. This yielded significant fits for the two age
closely matched levels of performance. In terms of the difference scores
groups for both Hexplicit (R2 > .84, p < .001) and Himplicit (R2 >
across sessions, the remaining groups still diverged, with the child group
.66, p < .01). The exponential coefficients (bs) were significantly
having a mean difference score of 13.7% (SD = 20.5), t(20) = 3.06, p =
different from zero (ts > 7.07, p < .001, for Hexplicit, and ts > 4.37,
.006, and the adult group having a mean difference score of −14.5% (SD
p < .001 for Himplicit), indicating significant learning within session
= 21.9), t(18) = −2.36, p = .030. Thus, although both these subgroups
C for both sequences for both groups. The power-law models for
were at equivalent levels of performance at the end of session A, the
the group-averaged data for the two types of Hebb sequence were
child group still increased their performance (i.e., they retained and per-
extrapolated to two additional trial points and the differences be-
haps consolidated their sequence knowledge), while performance of the
tween this extrapolation and each individual’s performance at the
adult group dropped during the delay.
one-year post-learning session (i.e., session D) was assessed (trial
Overall, there was no clear age-effect for Hexplicit, F(1, 53) = 2.5, p
by trial) and averaged (Figure 5). For Himplicit, both groups’ per-
= .12. Both groups showed retention although adults showed numer-
formance decreased across the delay (children, M difference score
ical but non-significant decrements (M difference score for children =
= −24.4%, SE = 3.6, t(27) = −6.82, p = .001; adults, M difference
2.1%, SE = 4.5, t(28) = .472, p = .640; M difference score for adults =
score = −35.7%, SE = 5.9, t(20) = −5.97, p = .001). Adults decreased
−8.1%, SE = 4.7, t(25) = −1.7, p = .097).
numerically, but not significantly, more than children (main effect of Age group, F(1, 47) = 3.4, p = .07, ɳ2p = .067). When the 75% of
3.2.2 | Across one week
the children that showed the highest score on trial 36 of session C (n = 21, M trial score = 90.0%, SD = 14.4) were compared with the
A similar power-law function was fitted to the averaged recall data of
75% of the adults that showed the lowest score on the end trial of
session B for the two types of Hebb sequence. This yielded significant
session C (n = 16, M trial score = 90.3%, SD = 16.6), adult perfor-
fits for the two age groups for both Hexplicit (R2 > .88, p < .001) and
mance dropped significantly more than that of children (Adults, M
Himplicit (R2 > .79, p < .001). The exponential coefficients (bs) were sig-
difference score = −46.6%, SD = 21.3; Children, M difference score
nificantly different from zero (ts > 8.18, p < .001, for Hexplicit, and ts >
= −20.0%, SD = 18.5, t(35) = 3.91, p < .001).
5.91, p < .001 for Himplicit), indicating significant learning within ses-
There was no age-effect, F < 1, for Hexplicit with both groups de-
sion B for both sequences for both groups. Again, the power-law mod-
creasing across the delay (adults, M difference score = −32.4%, SE =
els for the group-averaged data for the two types of Hebb sequence
5.9, t(20) = −5.4, p < .001; children, M difference score = −38.7%, SE =
were extrapolated to four additional trials and the differences between
5.1, t(27) = −7.63, p < .001).
this extrapolation and each individual’s performance at the one-week post-learning session (i.e., Session C) were assessed trial-by-trial and averaged (Figure 4). For Himplicit, planned comparisons revealed a main effect of Age group, F(1, 53) = 4.8, p = .033, ɳ2p = .08, with children
3.3 | Explicit knowledge of the sequences After session A, only 41.4% of the participants in the child group had
showing retention (M difference score = 3.9%, SE = 4.3, t(28) = .89, p =
noticed that there was another repetition in addition to the explicit
.38), in contrast with adults who showed a significant decrement across
sequence, while 73.1% of adults were aware of this. The χ2-test re-
the delay (M difference score = −9.9%, SE = 4.5, t(25) = −2.21, p = .037).
vealed a significant difference across age groups (χ2(1) = 5.6, p < .029).
Again, when the 75% of the children that showed the highest score on
By the end of session B, only 48.3% of the children had noticed the
trial 24 of session B (n = 21, M trial score = 83.1%, SD = 17.4) were com-
unannounced repetition while 84.6% of the adults were aware of it
pared with the 75% of the adults that showed the lowest score on the
(χ2(1) = 8.01, p < .005). Finally, by the end of session C, almost all par-
end trial of session B (n = 19, M trial score = 87.7%, SD = 20.3), children
ticipants had noticed the unannounced repetition (i.e., 79.3% in the
retained, and even significantly increased, performance across the delay
child group and 92.3% in the adult group), and the groups no longer
(M difference score = 12.0%, SD = 4.2, t(20) = 2.85, p = .010), while adult
differed significantly (χ2(1) = 1.86, p > .259). To further characterize
performance decreased numerically although not reliably (M difference
age differences in explicit knowledge of the sequence repetition for
score = −11.01%, SD = 5.3, t(18) = −2.077, p = .052).
these aware participants, we compared the time point at which the
There was no age-effect for Hexplicit, F < 1, with both groups
aware participants reported that they noticed the repetition and how
showing retention across the delay (child group, M difference score =
certain they were about the repetition (on a rating scale of 1 to 10).
−2.4%, SE = 4.3, t(28) = −.55, p = .59; adult group, mean = −3.4%, SE =
Overall, by the end of session C, all aware children had noticed the
3.3, t(25) = −1.01, p = .32).
repetition later (i.e., around the middle of the second session) in the task than aware adults (i.e., at the end of the first session), t(45) =
3.2.3 | Across one year
2.58, p < .013, and they were also less confident about their report (M = 8.3, SD = 1.8) than adults (M = 9.8, SD = .53), t(45) = −3.6, p < .001.
After one year, 28 children and 21 adults, that is, 89% of the
So, overall adults had gained more explicit knowledge of the sequence
original sample, returned to the lab. Only these participants were
and at a faster rate than children.
|
9 of 15
SMALLE et al. Hebb - Adults Extrapolated - Adults Hebb - Children Extraploated - Children
Explicit
Implicit 100
Correct responses (%)
Correct responses (%)
F I G U R E 4 Offline consolidation performance across one week as a function of Sequence Type (Explicit Hebb, Implicit Hebb) and Age group (Children and Adults). Hebb learning during session B is shown together with an extrapolation to four additional data points in session C. (1) Left panel: performance for the implicit Hebb sequence. (2) Right panel: performance for the explicit Hebb sequence. Error bars denote SEM
100 80 60 40 20 0
12
14
16
18
20
22
24
26
28
80 60 40 20 0 12
14
16
18
Trials
20
22
24
26
28
Trials
Session B
Session B
Session C
Session C
Hebb- Adults Extraploated - Adults Hebb - Children Extrapolated- Children
Explicit
100
Correct responses (%)
Correct responses (%)
Implicit F I G U R E 5 Offline consolidation performance across one week as a function of Sequence Type (Explicit Hebb, Implicit Hebb) and Age group (Children and Adults). Hebb learning during session C is shown together with an extrapolation to two additional data points in session D. (1) Left panel: performance for the implicit Hebb sequence. (2) Right panel: performance for the explicit Hebb sequence. Error bars denote SEM
80 60 40 20 0
24
26
28
30
32
34
36
38
100 80 60 40 20 0
24
26
28
Session C
30
32
34
36
38
Trials
Trials Session D
Session C
Session D
It is possible that the child–adult differences obtained for offline
Across the one-year delay, post-learning awareness further positively
consolidation of the implicit Hebb sequence (see main results) were
affected the retention scores for adults, but not for children, which is
driven by the fact that adults had gained more explicit knowledge of the
indicated by the significant interaction between Age and Awareness. We
repetitive occurrence of the unannounced repeating sequence. Contrary
further address this finding in the discussion.
to our expectation of an implicit-memory deficit for adults, this would imply an adult deficit in explicit memory for an unannounced repeating sequence that they had nonetheless noticed more than had the chil-
3.4 | Correlation with cognitive measures
dren. To investigate this, we added confidence ratings (as a measure of
We also aimed to find out whether explicit knowledge of the sequences
explicit knowledge or awareness of the sequence) as a moderator vari-
(i.e., confidence rating at the end of each session with inclusion of all par-
able into the univariate analysis of variance for the effect of Age-group
ticipants), vocabulary knowledge and verbal short-term memory capac-
on the retention of the unannounced repeating sequence (see also
ity are related to HRL in both groups. Therefore, a Pearson correlation
Supplementary Materials 1 for additional analyses of this issue, using
analysis was conducted on the magnitude of learning at the end of ses-
approximately matched subgroups of children and adults). A summary
sion A (Table 4). Magnitude of learning was defined as immediate recall
of the results is provided in Table 3. Overall, these results indicate supe-
on the final time point of the Hebb sequence, whilst controlling for the
rior delayed implicit memory for children across all offline periods, irre-
performance at the final time point for the filler sequence. In addition, we
spective of their level of awareness. The analysis also indicated a reliable
performed correlational analyses between our cognitive measures and
relationship between retention and the moderator Awareness so that
the extrapolated difference scores across each delay (see Tables 5–7). All
greater awareness of the unannounced repetition led to generally better
correlations show that HRL and subsequent offline retention were cor-
retention (an effect that was statistically reliable for the adults but not
related positively with post-learning awareness and digit span in adults,
for the children, see Tables 5–7 in the subsequent correlation analyses).
and with vocabulary scores in children.
|
SMALLE et al.
10 of 15
T A B L E 3 Results of the moderation analysis, with retention (i.e., the averaged difference between the extrapolated and the observed scores) of implicitly presented material as the dependent variable, Age group as the factor, and Awareness (i.e., confidence rating) of the unannounced sequence entered as a potential moderator Across four hours Variable
F
Across one week
Sig.
n
2
p
Across one year
F
Sig.
n
2
p
F
Sig.
n2p
Age group
9.59
.003**
.158
10.38
.002**
−.169
6.78
.012*
.131
Awareness
10.65
.002**
.173
11.33
.001**
.182
6.40
.015*
.124
.375
.015
2.38
.129
.045
6.03
.018*
.118
Age group*Awareness R
.800
2
Adjusted R2
.263
.256
.176
.219
.212
.121
2
R determination coefficient; *p < .05; **p < .01.
T A B L E 4 Partial correlations between the magnitude of HRL (while controlling for filler performance) and post-learning awareness of the unannounced sequence (collected after session A), Peabody vocabulary score, and digit span. Df equals 26 for children and 23 for adults. *p
